Risk factors for plastic bronchitis complicating severe Mycoplasma pneumoniae pneumonia in children and development of a machine learning-based predictive model

Abstract

Background

Plastic bronchitis (PB) is a rare but life-threatening complication of severe Mycoplasma pneumoniae pneumonia (SMPP) in children, and effective tools for its early prediction remain limited. This study aimed to identify risk factors for PB complicating SMPP and develop a machine learning-based predictive model.

Methods

This retrospective cohort study included 166 children (aged 1–14 years) with SMPP admitted between January 2022 and December 2024. Patients were classified into PB (n=51) and non-PB (n=115) groups based on bronchoscopy and randomly split into training and validation sets (7:3 ratio). Eight machine learning algorithms were developed and evaluated based on the area under the receiver operating characteristic curve (AUC), precision-recall curve (PRC), decision curve analysis (DCA), accuracy, and F1-score. The optimal model was selected and interpreted using SHapley Additive exPlanations (SHAP). Candidate risk factors were identified and further analyzed using multivariable logistic regression to develop a nomogram. The predictive model was assessed using both sets and further validated by bootstrap resampling.

Results

Generalized linear model (GLM) achieved the highest performance [AUC: 0.951, cross-validation AUC: 0.915, accuracy 0.880, F1-score: 0.825, area under the precision-recall curve (AUPRC): 0.856] in the training set, representing the optimal model. Fever duration, serum prealbumin, and D-dimer were the most influential features in SHAP analysis. Multivariable analysis confirmed these as independent risk factors: fever duration [odds ratio (OR): 1.550, 95% confidence interval (CI): 1.297–1.908, P<0.001], prealbumin (OR: 0.959, 95% CI: 0.939–0.976, P<0.001), and D-dimer (OR: 1.723, 95% CI: 1.157–2.627, P=0.009). The nomogram exhibited acceptable discriminative ability [C-index: 0.933 (95% CI: 0.887–0.980) in the training set; 0.831 (95% CI: 0.717–0.946) in the validation set; 0.913 (95% CI: 0.869–0.957) in bootstrap validation] and calibration (Brier score: 0.106, 0.142, and 0.103, respectively). The Hosmer-Lemeshow goodness-of-fit test yielded χ²=5.983 for internal validation [P=0.649 (>0.05)], suggesting acceptable model fit, with DCA confirming clinical utility across 0–90% threshold probabilities.

Conclusions

Fever duration, serum prealbumin, and D-dimer are independent risk factors for PB in children with SMPP. This internally derived nomogram achieved moderate discrimination and acceptable calibration in our limited single-center cohort. Further external validation in independent patient cohorts is therefore necessary to verify its generalizability and clinical utility.

Highlight box.

<bold>Key findings</bold>

• Machine-learning-driven analysis identified prolonged fever duration, decreased serum prealbumin and elevated D-dimer as independent risk factors for plastic bronchitis (PB) in children with severe Mycoplasma pneumoniae pneumonia (SMPP), and a three-variable nomogram demonstrated acceptable performance via internal hold-out and bootstrap validation.

<bold>What is known and what is new?</bold>

• SMPP in children may be complicated by PB, which can lead to severe respiratory sequelae. Flexible bronchoscopy is the gold-standard diagnostic test for PB but is too invasive for routine screening. Reliable non‑invasive predictive tools for PB remain limited.

• Using eight machine‑learning algorithms, we identified fever duration, prealbumin and D-dimer as three independent predictors for PB among children with SMPP. An internally validated nomogram was constructed based on these three readily available clinical and laboratory indicators.

<bold>What is the implication, and what should change now?</bold>

• The easy-to-use nomogram exhibits acceptable discrimination, calibration and net clinical benefit. It may assist clinicians in stratifying PB risk and reducing unnecessary bronchoscopy, pending further external validation.

Introduction

The outbreaks of Mycoplasma pneumoniae (MP), the most common pathogen responsible for lower respiratory tract infections in children, occur every 3–5 years (1-3). It accounts for 20–40% of all community-acquired pneumonia (CAP) cases. Mycoplasma pneumoniae pneumonia (MPP) typically presents as self-limiting persistent dry cough and fever (4). However, the emergence of macrolide-resistant MP strains since 2020 has increased the incidence of refractory MPP (RMPP), severe MPP (SMPP), and fulminant MPP (4,5). Compared with uncomplicated MPP, SMPP and RMPP are associated with substantially higher rates of complications (6,7). Specifically, the incidence of complications such as atelectasis, bronchiolitis obliterans, and bronchiectasis increases from 10.5% in uncomplicated MPP to 42% in SMPP and RMPP (8).

Respiratory tract infection represents the leading cause of plastic bronchitis (PB) in Asia, with MP being the predominant responsible pathogen (9-11). MP infection can induce severe airway inflammation, causing mucus plug formation and airway obstruction (12,13). Without prompt intervention, this can progress to bronchial casting, which is characterized by inflammatory bronchial stenosis, obstruction, and granulation tissue hyperplasia. This, in turn, can lead to long-term sequelae such as atelectasis, secondary bronchiectasis, and bronchiolitis obliterans, which significantly compromise pulmonary function and quality of life. Therefore, early diagnosis of PB and prompt bronchoscopic cast removal are critical to prevent life-threatening complications and reduce long-term morbidity (14,15).

Despite being essential for diagnosing and managing PB, the invasive nature and associated risks of flexible bronchoscopy limit its routine use (14,15). Consequently, an accurate, simple, non-invasive predictive model using readily available clinical parameters would assist clinical decision-making and reduce unnecessary bronchoscopic procedures. To address this need, we aimed to (I) identify risk factors for PB in children with SMPP using machine learning, (II) develop and internally validate a predictive model, and (III) construct a clinically applicable nomogram for early PB diagnosis. We present this article in accordance with the TRIPOD reporting checklist (available at https://tp.amegroups.com/article/view/10.21037/tp-2026-0579/rc).

Methods

Study design and participants

This retrospective cohort study included 166 children (aged 1–14 years) who were hospitalized in the Department of Respiratory Medicine at the Children’s Hospital of Shanghai Jiao Tong University School of Medicine with a discharge diagnosis of SMPP between January 2022 and December 2024. The patients were classified into PB (n=51) and non-PB (n=115) groups based on bronchoscopy and randomly split into training and validation sets (7:3 ratio) using a random number table (Table 1).

Table 1. Demographics and clinicopathologic characteristics of patients in the training and validation sets.

Demographics or characteristic	Training set			Validation set
	PB (n=36)	Non-PB (n=81)	P value	PB (n=15)	Non-PB (n=34)	P value
Gender			0.02			0.33
Male	23 (63.9)	34 (42.0)		7 (46.7)	11 (32.4)
Female	13 (36.1)	47 (58.0)		8 (53.3)	23 (67.6)
Age (months)	79.1±26.1	77.2±30.4	0.75	78.9±30.7	68.9±24.8	0.23
Fever duration (days)	11 [10–12]	8 [6–10]	<0.001	11.9±2.7	8.1±2.5	<0.001
Hypoxemia			<0.001			<0.001
Yes	11 (30.6)	4 (4.9)		8 (53.3)	2 (5.9)
No	25 (69.4)	77 (95.1)		7 (46.7)	32 (94.1)
Use of globulin			<0.001			0.007
Yes	10 (27.8)	2 (2.5)		6 (40.0)	2 (5.9)
No	26 (72.2)	79 (97.5)		9 (60.0)	32 (94.1)
Use of glucocorticoids			<0.001			0.02
Yes	33 (91.7)	36 (44.4)		13 (86.7)	18 (52.9)
No	3 (8.3)	45 (55.6)		2 (13.3)	16 (47.1)
Atelectasis (≥2 lung segments)			<0.001			0.004
Yes	16 (44.4)	11 (13.6)		10 (66.7)	8 (23.5)
No	20 (55.6)	70 (86.4)		5 (33.3)	26 (76.5)
Pleural effusion			0.02			0.72
Yes	13 (36.1)	14 (17.3)		3 (20.0)	10 (29.4)
No	23 (63.9)	67 (82.7)		12 (80.0)	24 (70.6)
MP DNA copies (104/mL)	6.5 [3.4–6.5]	5.0 [5.0–6.5]	0.77	6.5 [4.4–6.5]	6.5 [5.0–6.5]	0.65
MP 2063/2064			0.43			0.60
Yes	19 (52.8)	49 (60.5)		9 (60.0)	23 (67.6)
No	17 (47.2)	32 (39.5)		6 (40.0)	11 (32.4)
ESR (mm/h)	46 [30–54]	46 [37–59]	0.39	49.5±19.2	47.2±10.9	0.60
D-dimer (mg/L)	1.28 [0.93–2.97]	0.56 [0.35–0.95]	<0.001	1.1 [0.93–1.905]	0.78 [0.57–1.15]	0.055
CRP (mg/L)	21 [14–47.75]	9 [5–18]	<0.001	14 [5.5–33.5]	17.5 [7.5–32.75]	0.53
PCT (ng/mL)	0.14 [0.12–0.33]	0.09 [0.05–0.21]	0.02	0.31 [0.15–0.65]	0.12 [0.07–0.19]	0.02
ALT (U/L)	18 [18–29]	15 [12–23]	<0.001	18 [16.5–28]	14.5 [12–19]	0.03
Albumin (g/L)	35.6±4.8	41.4±3.5	<0.001	36.6±4.1	40.3±3.1	0.001
Prealbumin (mg/L)	80.5 [58.4–91.2]	105.8 [105.8–119]	<0.001	79.4 [61–113.1]	105.8 [105.8–112.1]	0.03
LDH (U/L)	384 [381–589]	344 [309–411]	<0.001	533.6±186.1	395.6±113.7	0.01
Ferritin (ng/mL)	157.9 [157.9–247.2]	143.7 [89.1–172]	<0.001	157.9 [122.1–157.9]	119.9 [95.4–157.9]	0.13
IgE (IU/mL)	133 [99.4–300.5]	133 [95.9–287]	0.79	133 [83.7–235]	133 [105.6–253.3]	0.55
BALF WBC (106/L)	1,416 [1,138–2,137]	1,250 [500–1,600]	0.01	780 [513–1,399]	1,350 [540–2,132]	0.61
BALF N (%)	68.5 [57.5–72]	70 [57–77]	0.22	68.5 [68.25–72]	73.5 [65.5–79.8]	0.15
BALF L (%)	18 [12.8–29.8]	14 [10–20]	0.11	16 [11–20.5]	14 [10.3–19.5]	0.42
BALF MP DNA copies (104/mL)	102 [25.8–731]	36.1 [2.7–158]	0.01	123.5 [66.2–277.3]	30.5 [8.2–101.9]	0.01
BALF MP 2063/2064			0.54			0.06
Yes	33 (91.7)	73 (90.1)		11 (73.3)	32 (94.1)
No	3 (8.3)	8 (9.9)		4 (26.7)	2 (5.9)
BALF IL-8 (102 pg/mL)	30.7 [17.9–42.3]	21.7 [9.8–38]	0.02	24.5 [16.9–47.8]	20.1 [10.8–31.4]	0.054
BALF IL-1β (pg/mL)	239.2 [102–503.2]	168.2 [49.7–498.2]	0.35	102.3 [55.5–368.6]	157.9 [69.7–560.4]	0.25
BALF IL-2 (pg/mL)	0.1 [0.1–0.1]	0.1 [0.1–0.1]	0.11	0.1 [0.1–1.3]	0.1 [0.1–0.1]	0.49
BALF IL-6 (pg/mL)	3.4 [0.1–132.7]	0.1 [0.1–53.5]	0.056	3.4 [0.1–97.9]	0.1 [0.1–30.3]	0.27
BALF IL-10 (pg/mL)	0.1 [0.1–0.1]	0.1 [0.1–0.1]	0.10	0.1 [0.1–2.4]	0.1 [0.1–5.3]	0.63
BALF TNF-α (pg/mL)	0.1 [0.1–0.1]	0.1 [0.1–0.1]	0.51	0.1 [0.1–0.1]	0.1 [0.1–0.1]	0.52

Data are presented as n (%), mean ± standard deviation, or median [interquartile range]. ALT, alanine aminotransferase; BALF, bronchoalveolar lavage fluid; CRP, C-reactive protein; DNA, deoxyribonucleic acid; ESR, erythrocyte sedimentation rate; IgE, immunoglobulin E; IL, interleukin; L, lymphocyte; LDH, lactate dehydrogenase; MP, Mycoplasma pneumoniae; n, number of patients; N, neutrophil; PB, plastic bronchitis; PCT, procalcitonin; TNF-α, tumor necrosis factor-alpha; WBC, white blood cell.

Diagnostic criteria

MP infection was diagnosed by either a positive serum-specific MP immunoglobulin M (IgM) antibody (acute-phase titer ≥1:160) or positive MP-DNA detected by real-time fluorescent polymerase chain reaction in throat swabs or bronchoalveolar lavage fluid (BALF). SMPP, diagnosed according to the Guidelines for the Diagnosis and Treatment of Mycoplasma pneumoniae Pneumonia in Children (2025 Edition) (16), was defined as severe CAP or MPP at risk of sequelae.

PB was diagnosed based on (I) bronchoscopic findings showing hyperemia and edema of the respiratory mucosa, increased mucus secretion, and bronchial luminal occlusion by inflammatory bronchial casts (BCs) that expanded into a branching plastic morphology when removed with biopsy forceps and immersed in normal saline (9 g/L); and (II) histopathological findings demonstrating BCs containing extensive inflammatory cells (predominantly eosinophils and neutrophils) and desquamated epithelial cells, with positive immunohistochemical staining for CD3, CD20, CD68, and myeloperoxidase (MPO). No blinding was performed for outcome assessment.

Inclusion and exclusion criteria

Inclusion criteria were (I) age ≤14 years; (II) SMPP diagnosis (16); (III) written informed consent obtained from parents or legal guardians; and (IV) complete clinical data.

Exclusion criteria comprised (I) recurrent respiratory tract infections, asthma, chronic pulmonary diseases, malignancies, hepatorenal diseases, connective tissue diseases, previous cardiac surgery, hematological disorders, or immunodeficiency disorders; (II) recovery phase of MPP; (III) concurrent infection with other pathogens or Mycobacterium tuberculosis; (IV) foreign body aspiration; or (V) incomplete clinical data.

Ethics statement

This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Children’s Hospital of Shanghai Jiao Tong University School of Medicine (approval No. 2026R084-E01). Written informed consent for the clinical procedures was obtained from the parents or legal guardians of all participants. For this retrospective study, the requirement for additional informed consent was waived by the Ethics Committee due to the use of anonymized data.

Data collection

The following variables were collected from all participants:

1. Demographic characteristics: sex and age.

2. Clinical manifestations: fever duration, length of hospitalization, duration of cough, chest pain, chest tightness, hypoxemia, and hemoptysis; pulmonary signs; and intra- and extrapulmonary complications.

3. Treatment history: macrolide therapy, oxygen therapy, intravenous immunoglobulin (IVIG), and glucocorticoids.

4. Laboratory and imaging parameters:

   • MP DNA copy numbers and 2063/2064 locus mutations;

   • Inflammatory markers: erythrocyte sedimentation rate (ESR), C-reactive protein (CRP), procalcitonin (PCT), and D-dimer;

   • Hepatic and nutritional markers: alanine aminotransferase (ALT), albumin, and prealbumin;

   • Lactate dehydrogenase (LDH) and ferritin;

   • BALF analysis: white blood cell (WBC) count, differential percentages, and interleukin (IL-2, IL-6, IL-8, IL-10) levels;

   • Imaging findings: presence of atelectasis and pleural effusion.

Statistical analysis and machine learning

Statistical analysis was performed using R software (version 4.3.2; R Foundation for Statistical Computing, Vienna, Austria). Thirty-one independent variables were processed to address potential multicollinearity using the following criteria: variance inflation factor (VIF) threshold ≤5, high correlation cutoff ≤0.7, and standard deviation threshold for constant variables ≤0.05. Eight machine learning algorithms—random forest (RF), support vector machine (SVM), generalized linear model (GLM), extreme gradient boosting (XGBoost), gradient boosting machine (GBM), k-nearest neighbors (KNN), decision tree (DT), and naive Bayes (NB)—were developed and evaluated in the training and validation sets. In the preprocessing workflow, feature normalization followed our consistent training-data-only processing strategy used for collinearity and variable screening. Z-score scaling parameters were estimated solely from the training set and reused on the validation cohort to avoid information leakage, while tree algorithms required no scaling.

All preprocessing and feature screening procedures were fitted exclusively on the training dataset to avoid data leakage. Missing value imputation, median-based outlier truncation, one-hot encoding and standardization were trained using only training samples; the calculated median, quantile and scaling parameters were directly applied to the validation set without re-estimation. Variable filtering including variance screening, VIF and Pearson correlation filtering was performed within training data only, and no statistics derived from validation samples leaked into the modeling pipeline.

To mitigate unstable performance evaluation stemming from the small sample size and limited positive PB events, we implemented five repeats of stratified 10-fold cross-validation (CV) exclusively within the 70% training cohort. The 30% hold-out validation set was reserved only for auxiliary external reference and not involved in CV or model screening. We trained eight machine learning algorithms for exploratory performance comparison, CV metrics serve as the primary screening criterion, and hold-out validation results are only used for secondary supplementary reference, and the final interpretable clinical nomogram was built using parsimonious L2-regularized GLM with superior anti-overfitting performance, instead of complex ensemble models.

Descriptive statistics and group comparisons

Categorical variables were summarized as counts (percentages) and compared using the χ² test or Fisher’s exact test, as appropriate. Continuous variables with normal distribution, expressed as mean ± standard deviation, were compared using Student’s t-test. Non-normally distributed continuous variables, presented as median [interquartile range (IQR)], were compared using the Mann-Whitney U test.

Predictive model development

A predictive model for PB was constructed using the rms package in R. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), precision-recall curve (PRC), decision curve analysis (DCA), accuracy, and F1-score. The optimal model was selected based on the above-mentioned metrics and interpreted using SHapley Additive exPlanations (SHAP) to identify the most influential features. Candidate risk factors were subsequently analyzed using multivariable logistic regression to develop a nomogram. Then, internal validation was performed using bootstrap resampling (1,000 iterations) in addition to the hold-out validation set. Afterward, model calibration was evaluated using calibration curves and the Brier score. No differences in study setting, eligibility criteria, outcome definition or predictor variables existed between the validation cohort and development cohort. Statistical significance was set at two-tailed P<0.05.

Results

Baseline characteristics

Among the 166 children with SMPP, 51 (30.7%) developed PB, while 115 (69.3%) did not (Figure S1). Table 1 summarizes the baseline characteristics. Compared with the non-PB group, the PB group had longer fever duration {median [IQR]: 11 [10–13] vs. 8 [6–10] days; P<0.001}, higher rates of hypoxemia (37.3% vs. 5.2%; P<0.001), glucocorticoid use (90.2% vs. 47.0%; P<0.001), IVIG administration (31.4% vs. 3.5%; P<0.001), and multilobar atelectasis (≥2 lobes; 51.0% vs. 16.5%; P<0.001). Laboratory markers showed significantly higher D-dimer, CRP, PCT, ALT, LDH, ferritin, IL-6, and IL-8 levels and lower albumin and prealbumin levels in BALF in the PB group (all P<0.05). Age, sex, pleural effusion, MP 2063/2064 mutation rate, ESR, immunoglobulin E (IgE), BALF WBC count and differential count, or IL-1β, IL-2, IL-10, and tumor necrosis factor-alpha (TNF-α) levels showed no significant differences (Table S1).

Machine learning model comparison

Following sequential filtering for constant predictors (standard deviation threshold ≤0.05) and multicollinearity (VIF and Pearson correlation screening), 31 independent variables were retained to construct machine learning prediction models (Figures S2-S4). All screening thresholds were generated from training data to prevent leakage-induced overoptimistic indices. Given its superior performance, GLM was selected for further analysis. The exceptional training performance [AUC: 0.951, CV-AUC: 0.915, area under the precision-recall curve (AUPRC): 0.856, CV-AUPRC: 0.730, F1 score: 0.825, CV-F1 score: 0.584] (Table 2, Figure S5) suggests potential overfitting; however, validation set performance (AUC: 0.900) supported model robustness.

Table 2. Performance metrics of eight machine learning models (cross-validation vs. training set) for plastic bronchitis prediction.

Metrics	GLM	KNN	SVM	RF	GBM	XGBoost	DT	NB
CV-AUC	0.915	0.905	0.895	0.797	0.904	0.896	0.794	0.894
CV-AUPRC	0.730	0.691	0.728	0.538	0.670	0.698	0.250	0.670
CV-F1 score	0.584	0.564	0.738	0.507	0.491	0.635	0.636	0.738
AUC	0.951	0.932	0.959	0.919	0.939	0.950	0.871	0.918
AUPRC	0.856	0.804	0.881	0.784	0.804	0.860	0.622	0.834
Accuracy	0.880	0.872	0.880	0.872	0.889	0.889	0.855	0.863
Sensitivity	0.917	0.861	0.944	0.944	0.861	0.944	0.889	0.944
Specificity	0.864	0.877	0.852	0.840	0.901	0.864	0.840	0.827
Precision	0.750	0.756	0.739	0.723	0.795	0.756	0.711	0.708
F1 Score	0.825	0.805	0.829	0.819	0.827	0.840	0.790	0.810

AUC, area under the curve; AUPRC, area under the precision-recall curve; CV-AUC, cross-validation area under the receiver operating characteristic curve; CV-AUPRC, cross-validation area under the precision-recall curve; CV-F1, cross-validation F1; DT, decision tree; GBM, gradient boosting machine; GLM, generalized linear model; KNN, k-nearest neighbors; NB, naive Bayes; RF, random forest; SVM, support vector machine; XGBoost, extreme gradient boosting.

GLM model interpretation

The GLM model identified the top 10 predictive features from 31 candidate variables: fever duration, serum prealbumin, serum D-dimer, BALF IL-6, serum albumin, BALF neutrophil percentage, sex, BALF IL-8, MP DNA copy number, and serum PCT (Figures 1,2). SHAP analysis revealed that fever duration, prealbumin, and D-dimer were the three most influential features. Partial dependence plots (Figure 3) showed that fever duration and D-dimer were positively associated with PB probability, whereas prealbumin demonstrated a negative association. SHAP summary plots (Figure 4) further illustrated the direction and magnitude of these associations.

Figure 1.

SHAP variable importance plot for the top 10 features in the optimal GLM model. The horizontal axis shows the mean absolute SHAP value, representing the overall contribution of each feature to the predictive model. Features are ranked in a descending order of importance. Fever duration, PA, and D-dimer were the three most influential features in predicting the outcome. BALF, bronchoalveolar lavage fluid; DNA, deoxyribonucleic acid; GLM, generalized linear model; IL, interleukin; MP, Mycoplasma pneumoniae; PA, prealbumin; PCT, procalcitonin; SHAP, SHapley Additive exPlanations.

Figure 2.

SHAP dependence plot of the top 10 features in the optimal GLM model. The plot illustrates the direction and magnitude of each feature’s contribution to the model prediction, with feature values shown on the horizontal axis and corresponding SHAP values presented on the vertical axis. BALF, bronchoalveolar lavage fluid; DNA, deoxyribonucleic acid; GLM, generalized linear model; IL, interleukin; MP, Mycoplasma pneumoniae; PCT, procalcitonin; SHAP, SHapley Additive exPlanations.

Figure 3.

PDPs for the three key predictors of PB. PDPs illustrating the marginal effect of (A) fever duration (days), (B) prealbumin level, and (C) D-dimer level on the mean predicted probability of PB. PB, plastic bronchitis; PDP, partial dependence plot.

Figure 4.

SHAP value versus feature value scatter plots for the three key predictors of PB. Scatter plots showing the relationship between feature value and corresponding SHAP value for (A) fever duration (days), (B) prealbumin, and (C) D-dimer. Points are colored by PB status (blue = non-PB, red = PB). PB, plastic bronchitis; SHAP, SHapley Additive exPlanations.

Model calibration and risk stratification

Decile calibration analysis demonstrated acceptable agreement between predicted and observed PB probabilities across all risk strata (Figure 5). The monotonic increase in observed PB prevalence from the 1st to 10th decile confirmed moderate risk stratification capacity.

Figure 5.

Decile calibration plot of the predicted probability vs. actual rate of PB. Patients were stratified into 10 risk deciles (1 = lowest risk, 10 = highest risk) according to the predicted probability. The blue line shows the mean predicted probability, and the red line demonstrate the actual observed PB rate in each decile. Close alignment between predicted and actual values indicates acceptable model calibration and risk stratification ability. PB, plastic bronchitis.

Predictive model development and validation

Multivariable logistic regression

The three most influential features (fever duration, prealbumin, and D-dimer) were entered into a multivariable logistic regression model:

$$\text{Riskscore}=-1.81+0.438\times \text{feverduration(days)}-0.042\times \text{prealbumin(mg/L)}+0.544\times \text{D-dimer(mg/L)}$$ [1]

Then, a nomogram was constructed for clinical use (Table 3, Figure 6).

Table 3. Univariate and multivariate logistic regression analyses of risk factors for PB in children with SMPP.

Variable	Univariate analysis				Multivariate analysis
	Coef.	Wald	OR (95% CI)	P	Coef.	Wald	OR (95% CI)	P
Intercept	–	–	–	–	−1.81	1.816	0.164 (0.01–2.26)	0.17
Fever duration (days)	0.451	30.150	1.571 (1.35–1.87)	<0.001	0.438	20.224	1.55 (1.30–1.91)	<0.001
PA (mg/L)	−0.047	26.703	0.954 (0.94–0.97)	<0.001	−0.042	18.151	0.959 (0.94–0.98)	<0.001
D-dimer (mg/L)	0.616	16.624	1.851 (1.39–2.53)	<0.001	0.544	6.892	1.723 (1.16–2.63)	0.009

CI, confidence interval; Coef., regression coefficient; OR, odds ratio; PA, prealbumin; PB, plastic bronchitis; SMPP, severe Mycoplasma pneumoniae pneumonia.

Figure 6.

The nomogram for predicting the Risk of PB in children with SMPP. A clinical prediction tool that assigns points for each independent risk factor (fever duration, prealbumin, and D-dimer). Total points are summed to calculate the individual probability of PB. PB, plastic bronchitis; SMPP, severe Mycoplasma pneumoniae pneumonia.

Model performance

The nomogram demonstrated acceptable discrimination [C-index: 0.933 (95% confidence interval (CI): 0.887–0.980) in the training set; 0.831 (95% CI: 0.717–0.946) in the validation set; 0.913 (95% CI: 0.869–0.957) in bootstrap validation] and good calibration (Brier score: 0.106, 0.142, and 0.103, respectively) (Figures 7,8). The slight performance decrease from the training to the validation set (maximum ΔC-index: 0.102; maximum ΔBrier score: 0.039) indicates minimal overfitting. Additional performance metrics across the three sets were as follows: 0.846 in the training set, 0.755 in the validation set, and 0.819 in bootstrap validation for accuracy; 0.889 in the training set, 0.800 in the validation set, and 0.863 in bootstrap validation for sensitivity; 0.827 in the training set, 0.735 in the validation set, and 0.800 in bootstrap validation for specificity; and 0.781 in the training set, 0.667 in the validation set, and 0.746 in bootstrap validation for F1-score (Figures 9-11).

Figure 7.

ROC curves and AUC values of the nomogram in training, validation, and bootstrap sets. ROC curves evaluating the discriminative ability of the nomogram in the training set (red), validation set (blue), and bootstrap validation set (green). Corresponding AUC values with 95% CI are presented in the inset. AUC, area under the curve; CI, confidence interval; ROC, receiver operating characteristic.

Figure 8.

Calibration curves and Brier scores of the nomogram. Calibration curves assessing agreement between predicted and observed probabilities of PB in the training (red), validation (blue), and bootstrap validation sets (green). The dashed diagonal line represents ideal calibration. Brier scores (0–1; lower values indicate better calibration) are shown for each dataset. PB, plastic bronchitis.

Figure 9.

Confusion matrix heatmaps of the nomogram. PB, plastic bronchitis. Confusion matrix heatmaps illustrating the predictive performance of the nomogram in (A) training set, (B) validation set, and (C) bootstrap validation set. Colors represent the frequency of each classification (true negative, false positive, false negative, and true positive). PB, plastic bronchitis.

Figure 10.

Quantitative performance metrics of the nomogram. Bar plots comparing the accuracy, AUC, F1-score, sensitivity, and specificity of the nomogram in the training set (red), validation set (blue), and bootstrap validation set (green). AUC, area under the curve.

Figure 11.

Distribution of predicted probabilities in training, validation, and bootstrap sets. Density plots showing the distribution of predicted probabilities of PB for non-PB (pink) and PB (blue) cases in the (A) training set, (B) validation set, and (C) bootstrap validation set. PB, plastic bronchitis.

Clinical DCA

DCA demonstrated net clinical benefit of the nomogram across threshold probabilities of 0–90%, with stable performance in both training and validation sets (Figure 12). The model outperformed both “treat-all” and “treat-none” strategies.

Figure 12.

Decision curve analysis of the nomogram. Decision curve analysis evaluating the clinical utility of the nomogram across a range of threshold probabilities. The red line represents the nomogram, the blue dashed line represents the “all intervention” strategy, and the gray solid line represents the “no intervention” strategy.

Individual risk factor performance

As standalone predictors, fever duration showed the highest discriminatory ability [C-index: 0.809 (95% CI: 0.746–0.872)], followed by D-dimer [C-index: 0.779 (95% CI: 0.710–0.848)] and prealbumin [C-index: 0.771 (95% CI: 0.680–0.861)]. Optimal cutoff values were 8.5 days for fever duration, 93.75 mg/L for prealbumin, and 0.925 mg/L for D-dimer (Tables 4,5, Figure 13).

Table 4. Associations of elevated D-dimer, prolonged fever duration, and decreased prealbumin with PB in children with SMPP.

Risk factor	PB, n (%)	Non-PB, n (%)	OR (95% CI)	P
Fever duration (>8.5 days)	48 (94.1)	46 (40.0)	24.0 (7.05, 81.67)	<0.001
PA (<93.75 mg/L)	36 (70.6)	16 (13.9.1)	14.85 (6.67, 33.08)	<0.001
D-dimer (>0.925 mg/L)	45 (88.2)	39 (33.9)	14.6 (5.74, 37.24)	<0.001

CI, confidence interval; n, number of patients; OR, odds ratio; PA, prealbumin; PB, plastic bronchitis; SMPP, severe Mycoplasma pneumoniae pneumonia.

Table 5. Diagnostic performance of fever duration, prealbumin and D-dimer for predicting PB in children with SMPP.

Diagnostic test evaluation indices	Fever duration (>8.5 days)	PA (<93.75 mg/L)	D-dimer (>0.925 mg/L)
AUC (95% CI)	0.809 (0.746, 0.872)	0.771 (0.680, 0.861)	0.779 (0.710, 0.848)
SEN (95% CI)	0.941 (0.877, 1.0)	0.706 (0.581, 0.831)	0.882 (0.794, 0.971)
SPE (95% CI)	0.60 (0.51, 0.69)	0.861 (0.798, 0.924)	0.661 (0.574, 0.747)
PPV (95% CI)	0.511 (0.410, 0.612)	0.868 (0.806, 0.930)	0.536 (0.429, 0.642)
NPV (95% CI)	0.958 (0.912, 1.004)	0.692 (0.567, 0.818)	0.927 (0.870, 0.983)
PLR (95% CI)	2.353 (1.862, 2.974)	2.927 (1.901, 4.506)	2.602 (1.978, 3.422)
NLR (95% CI)	0.098 (0.032, 0.297)	0.197 (0.121, 0.321)	0.178 (0.083, 0.382)

AUC, area under the curve; CI, confidence interval; NLR, negative likelihood ratio; NPV, negative predictive value; PA, prealbumin; PB, plastic bronchitis; PLR, positive likelihood ratio; PPV, positive predictive value; SEN, sensitivity; SMPP, severe Mycoplasma pneumoniae pneumonia; SPE, specificity.

Figure 13.

ROC curves of fever duration, prealbumin, and D-dimer for predicting PB. (A) ROC curve of the fever duration for predicting PB in children with SMPP. The optimal cutoff value was 8.5 days, with a sensitivity of 0.941 and a specificity of 0.600. The AUC was 0.809 (95% CI: 0.746–0.872). (B) ROC curve of the prealbumin for predicting PB in children with SMPP. The optimal cutoff value was 93.75 mg/L, with a sensitivity of 0.706 and a specificity of 0.861. The AUC was 0.771 (95% CI: 0.680–0.861). (C) ROC curve of D-dimer for predicting PB in children with SMPP. The optimal cutoff value was 0.925 mg/L, with a sensitivity of 0.882 and a specificity of 0.661. The AUC was 0.779 (95% CI: 0.710–0.848). AUC, area under the curve; CI, confidence interval; PB, plastic bronchitis; ROC, receiver operating characteristic; SMPP, severe Mycoplasma pneumoniae pneumonia.

Discussion

MPP is the leading cause of CAP in Chinese children aged ≥5 years. Direct MP-mediated cytotoxicity and dysregulated host immune responses increase disease severity, with SMPP and RMPP conferring substantial risks of intrapulmonary and extrapulmonary complications (16). PB, a critical complication of SMPP, is associated with significant morbidity and can be fatal. Early identification of PB risk factors is essential for improving clinical outcomes (16).

Key findings and pathophysiological mechanisms

We strictly separated training and validation data during all preprocessing and variable selection steps to avoid information leakage, a common factor causing artificially high predictive performance. Every statistical rule for feature processing was trained on training samples only, which enhanced the credibility of validation-set metrics. This study identified fever duration, serum prealbumin, and D-dimer as independent predictors of PB in children with SMPP. The derived nomogram demonstrated acceptable discrimination (C-index: 0.933) and clinical utility (DCA: 0–90%) using routinely available parameters.

SHAP and DCA were applied to interpret nonlinear risk associations and net clinical benefit, which cannot be captured by conventional simple logistic regression. Multiple machine learning models served only as auxiliary comparisons rather than core clinical prediction models. We performed multi-algorithm screening to validate the stability of the three risk predictors, instead of emphasizing complex machine learning architectures.

Fever duration reflects persistent MP infection and host inflammatory response. Our optimal cutoff (8.5 days) is shorter than the 10 days reported by Gong et al. for RMPP (17), suggesting that PB may precede refractory disease. This aligns with Zhao et al.’s study (18), which demonstrated prolonged fever in PB-complicated RMPP, as well as Huang et al.’s research (19), which identified fever duration as a risk factor for mucus plug formation in adenovirus pneumonia.

Prealbumin, a negative acute-phase reactant with a short half-life (~2 days), has superior sensitivity, compared to albumin, for detecting nutritional compromise and inflammation severity (20). Our cutoff (<93.75 mg/L) is substantially lower than Wang et al.’s 141.6 mg/L for SMPP (21) and Zhang et al.’s 144.5 mg/L for bronchial mucus plugs (14), likely reflecting the more severe inflammatory burden of PB. The inverse relationship between prealbumin and IL-6 (20) supports its role as a proxy for cytokine-driven pathology.

D-dimer elevation indicates fibrinolytic activation secondary to vascular endothelial injury and hypoxia. Our findings corroborate Qiu et al. (22) and Zheng et al.’s results (23), which demonstrated the superiority of D-dimer over conventional markers (CRP and LDH) for predicting MPP severity. Furthermore, the negative correlation with prealbumin suggests the coupled activation of inflammatory and coagulation cascades (Table 6).

Table 6. Comparison with existing predictive models.

Study	Sample size	Risk factors	Methods	Validation	Key limitations
Ma et al. (24)	126	MP-DNA, pulmonary atelectasis, D-dimer	Multivariate logistic regression analysis	None	MP-DNA delay (1–3 days); no cutoff values
Zhang et al. (25)	120	Fever, extrapulmonary complications, pleural effusion, cough duration, LDH	Multivariate logistic regression analysis	None	No pathogen exclusion; no cutoff values
Zhang et al. (14)	255	Prealbumin, corticosteroid timing, CRP, LDH	Multivariate logistic regression analysis	None	No pathogen exclusion; treatment variables (subjective)
Zhao et al. (18)	547	Peak body temperature, neutrophil ratio (N%), PLT, IL-6, LDH and pulmonary atelectasis	LASSO + multivariate logistic regression analysis	External	Numerous indicators; complex; no pathogen exclusion; no cutoff values; cytokine delay (1–2 days); single-center
Xu et al. (26)	1,271	Age, intrapulmonary complications, multiple cytokines	Multivariate logistic regression analysis	External	Cytokine delay; complex; no pathogen exclusion; no cutoff values; single-center
Current study	166	Fever, prealbumin, D-dimer	ML + multivariate logistic regression analysis	Train-test split validation + bootstrap resampling	Single-center; retrospective

CRP, C-reactive protein; DNA, deoxyribonucleic acid; IL-6, interleukin-6; LASSO, least absolute shrinkage and selection operator; LDH, lactate dehydrogenase; ML, machine learning; MP, Mycoplasma pneumoniae; PLT, platelet count.

In addition, Xu et al. (27) constructed a 5-variable PB nomogram including age, fever peak, atelectasis, albumin and LDH, with internal AUC 0.8768 and external AUC 0.8399. Our model only adopts three routine non-imaging indicators and delivers comparable predictive efficiency for easier bedside use. Besides, we adopted eight machine learning algorithms and SHAP for objective feature screening, which was absent in their study. Our model offers three advantages: (I) objective, routinely available parameters eliminating treatment-related subjectivity; (II) rapid turnaround (all tests available within hours); (III) stability of the results since multicollinearity was assessed among all candidate variables before model construction; and (IV) machine learning–optimized integration with demonstrated calibration and clinical utility.

Limitations

This study has several limitations. First, our cohort of 16 children with SMPP was relatively small compared with existing PB prediction models, and the single-center retrospective design inevitably introduced selection bias. A simple 7:3 train-test split may yield unstable predictive metrics given our limited PB events; thus, we conducted five repeats of stratified 10-fold CV on training data to generate more robust performance values, and we avoided overinterpreting optimistic training-set results throughout the Results section. Second, immune function (lymphocyte subsets and immunoglobulins) and pulmonary function tests were not assessed, and these may refine risk stratification. Third, external validation in independent cohorts, which was not performed, is essential before clinical implementation. Fourth, the obvious gap between training and cross-validated F1-scores is attributed to the limited number of PB positive events in the small training cohort, indicating mild overfitting risk.

Conclusions

Fever duration, serum prealbumin, and D-dimer are independent, readily available predictors of PB in children with SMPP. The internally validated nomogram demonstrates acceptable discrimination and calibration, supporting its potential for clinical decision-making, yet pending external validation. Children with fever lasting >8.5 days, prealbumin <93.75 mg/L, and D-dimer >0.925 mg/L warrant heightened surveillance for PB and early bronchoscopy.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Ethics Committee of Children’s Hospital of Shanghai Jiao Tong University School of Medicine (approval No. 2026R084-E01). The requirement for individual patient informed consent was waived due to the retrospective nature of the study.

Data Sharing Statement

Available at https://tp.amegroups.com/article/view/10.21037/tp-2026-0579/dss

tp-15-9-365-dss.pdf ^{(74.5KB, pdf)}

DOI: 10.21037/tp-2026-0579